Existing memories may be classified into volatile working memories (such as static random access memories (SRAMs) and dynamic random access memories (DRAMs)), and nonvolatile storage memories (such as NAND flash memories and hard disk drives (HDDs)). The volatile memories consume large energy due to leakage current caused in SRAMs, and refresh current caused in DRAMs.
In order to solve this problem, various nonvolatile memories have been considered as working memories replacing SRAMs and DRAMs.
The working memories, however, are accessed more often in an active state than in a standby state. Since a large writing charge (Qw) is required in the active state, energy required for a write operation increases. As a result, the energy saved during the standby state due to their nonvolatile characteristic is completely consumed in the active state, and therefore the total energy consumption is difficult to be decreased. This is called “nonvolatile memories' historical dilemma.” No existing product has solved this problem.
Recently simulation using state-the-art MTJ performance was performed. The simulation result showed reduction in energy consumption if a STT (Spin Transfer Torque)-MRAM (Magnetic Random Access Memory) is used as a lowest level cache memory (LLC (Last Level Cache)).
If the STT-MRAM is used as a cache memory in a layer upper than the LLC, the frequency at which it is accessed considerably increases. Therefore, considerable energy is consumed. Therefore, the aforementioned energy consumption problem cannot be solved.